應用生成資料集增強遙測影像之道路破壞辨識

Abstract

災害發生後，快速準確地評估道路損壞位置及影響程度對於及時災害應變至關 重要。傳統的評估方法仰賴於民眾回報和勘災人員現場調查。然而，在大範圍災害 中，這些方法往往耗時且費力，造成後續物資配給、救援工作及民眾疏散的延誤。 因此，本研究將透過應用遙測影像以及深度學習技術，進行大範圍快速地道路災損 辨識。針對目前應用遙測影像於道路損壞辨識任務的相關研究中，因災後影像中破 壞道路數量稀少及缺乏災前後成對影像作為訓練資料集導致深度學習模型辨識能 力不佳的狀況，本研究利用生成資料集進行資料擴增，來提升卷積神經網路(CNNs) 的效能，使用高斯噪點和生成對抗網路(GANs)等方法，由災前影像產生共四種 具有不同破壞特徵的生成災後影像。此方法增加了影像中破壞類別數量，並提供成 對的災前後影像作為訓練資料集，有助於模型辨識破壞道路的能力，並提升模型在 不同災害情境中的適應能力。此外，本研究提出了一個結合道路提取模型和孿生模 型的道路破壞辨識架構，並將輸出的分類結果進一步投影至地圖，進行路段破壞程 度分級。本研究結果顯示，有使用生成資料進行訓練的模型在道路破壞辨識能力上 有顯著提升，特別是透過生成對抗網路生成之有當地破壞特徵的資料集提升最多。 未來的研究將集中於在不同的光線條件、色調和地形下生成和評估合成損壞特徵， 以進一步提升損壞識別模型的穩健性。

Assessing road damage in the aftermath of disasters is crucial for timely and effective emergency response. Traditional assessment methods, which rely on civilian reports and on-site surveys, are often time-consuming and labor-intensive, especially during widespread disasters. This research addresses these limitations by applying remote sensing imagery and deep learning techniques for large-scale, rapid road damage identification. Existing studies on road damage identification using remote sensing imagery face challenges due to the scarcity of damaged road data and the lack of paired pre- and post-disaster images for training datasets. To overcome these issues, this study utilizes synthetic data for data augmentation to enhance the performance of Convolutional Neural Networks (CNNs). Using methods such as Gaussian noise and Generative Adversarial Networks (GANs), we generate four types of synthetic post-disaster images from pre-disaster images, each reflecting different damage characteristics. This approach increases the variety of damage categories in the images and provides paired pre- and post-disaster images for training datasets, improving the model's ability to identify damaged roads and its adaptability to different disaster scenarios. Additionally, we propose a road damage recognition framework that combines the Road Extraction Model and the Siamese damage identification model. The output prediction masks are projected onto a map, and road segments are classified to determine the extent of damage. Our results show significant improvements in the models' road damage identification capabilities when trained with synthetic data, particularly with datasets generated using GANs that include local damage features. Future research will focus on generating and evaluating synthetic damage features under various lighting conditions, color tones, and terrains to further enhance the robustness of damage identification models.